论文阅读学习 - 深度学习网络模型分析对比

深度学习网络模型分析对比

[Paper - An Analysis of Deep Neural Network Models for Practiacal Applications]

从准确率Accuracy、内存占用Memory Footprint、参数量Parameters、计算量(运算次数)Operations Count、推断时间Inference Time 和 功耗Power Consumption 几个方面对比 2016年前的模型表现.

一些结论：

• 功耗与 BatchSize 和 网络结构无关 - power consumption is independent of batchsize and architecture.
• 准确率和推断时间存在双曲线关系 hyperbolic relationship - accuracy and inference time are in a hyperbolic relationship.
• 能量约束是最大可达准确率和模型复杂度的上界 - energy constraint is an upper bound on the maximum achievable accuracy and model complexity.
• 计算量(运算次数)是推断时间的可靠估计 - the number of operations is a reliable estimate of the inference time.

平台与评价标准：

• Top-1 准确率，single central-crop sampling technique
• 推断时间和内存占用，Torch7，cuDNN-v5，CUDA-v8.
• 基于 JetPack-2.3 NVIDIA Jetson TX1 board (nVIDIA)：64-bit ARM R A57 CPU, a 1 T-Flop/s 256-core NVIDIA Maxwell GPU and 4 GB LPDDR4 of shared RAM.
• 运算次数基于开源平台 - torch-opcounter
• 功耗，

实现结果：